llm-tool-call-grpo-lora-Qwen1.5B

🛠️ LLM Tool Call Fine-Tuning — SFT + GRPO

Fine-tuning Qwen 2.5-1.5B to make structured JSON tool calls using SFT and Reinforcement Learning (GRPO).

---

🧪 Experiment

Train a small LLM to respond with structured tool calls instead of plain text:

{"name": "get_weather", "arguments": {"location": "Paris"}}

Three steps:

1. SFT — teach the model using 500 real examples from glaive-function-calling-v2
2. GRPO — improve using reward functions (no labeled answers needed)
3. Eval — compare SFT vs GRPO side by side on 12 test queries

---

📊 Results

Key finding: SFT model answered questions directly in plain text (never used tools). GRPO model learned to always respond with structured JSON tool calls. Tested on 12 queries across weather, calculator, search, stocks, translation and unit conversion.

---

🚀 Run on Colab

!pip install -q transformers datasets peft trl accelerate bitsandbytes
!python EXP_STEP1_sft.py    # ~20 min
!python EXP_STEP2_grpo.py   # ~30 min
!python EXP_STEP3_compare.py

---

📁 Files

---

🔧 Stack

Qwen2.5-1.5B • QLoRA • GRPO • glaive-function-calling-v2 • trl • peft

---

📎 References

• glaive-function-calling-v2
• GRPO — DeepSeekMath
• QLoRA paper